Technology does not destroy professions in bulk. It dismantles positions task by task.
This is the most counterintuitive finding of recent work on automation: a significant share of French workers currently occupy positions where a substantial proportion of their tasks are completely automatable by AI agents. Right now, with the capabilities currently available.
This observation deserves examination for what it does not say. It does not say these jobs will disappear. It says that the actual work of these people is changing in nature, that certain tasks they perform today can be taken over by automated systems, and that businesses, unions, and public authorities have not yet decided what to do about it.
It is this collective decision, more than the technology itself, that will determine whether agentic AI works to the advantage of workers or at their expense.
The Essentials
Recent work, notably that of Axelle Arquié at the CEPII as part of the Observatory of Threatened and Emerging Jobs, estimates that a significant share of French workers occupy positions where a substantial portion of their tasks is completely automatable by AI agents. The task-based approach reveals that exposure to AI varies according to the precise composition of positions, not solely according to socioprofessional categories. Only one-third of large French companies have launched agentic AI projects. The question that structures the debate is political: who decides the pace of job recomposition, according to what rules, and with what compensation for those whose work transforms fastest?
The Task-Based Approach Changes Everything in the Diagnosis
For a long time, the debate on AI and employment was organized around lists of professions. One consulting firm would publish a ranking of the most threatened occupations: accountants, lawyers, secretaries leading the list. Another study would respond that manual jobs would resist better. This framework was misleading because it treated jobs as homogeneous blocks.
The reality is more granular. A corporate lawyer position includes very different tasks: documentary research and jurisprudence synthesis, drafting standard contracts, negotiation, strategic advice, relationship management. The first are highly automatable today. The last are much less so. A profession-by-profession analysis tool would say that “the lawyer is exposed to 60%.” A task-based analysis would say that 40% of time spent on documentary tasks can be handled by an agent, which either liberates or eliminates, depending on the employer’s choice, those 40% of their time.
It is this distinction that the research of Axelle Arquié, economist at CEPII, places at the center of the analysis. The approach breaks down positions into elementary tasks, assesses the automatability of each with current AI agent capabilities, then scales back up to the position level. The result is a diagnostic tool considerably more precise than profession-based approaches, and far more useful for thinking about adaptation policies.
This methodological shift displaces the question from “which profession will survive” to “what job recomposition is possible, and who decides it.” It is a shift from technical terrain to political terrain.
What Exposure to AI Really Reveals
The estimate of a significant share of highly exposed workers must be read carefully. It identifies a threshold of significant exposure, a substantial proportion of completely automatable tasks, and observes that a non-negligible fraction of the working population already exceeds this threshold. But exposure is not uniformly distributed.
The most exposed positions concentrate in support functions of large organizations: administrative management, repetitive data analysis, document processing, level 1 customer service. These are often intermediate positions, neither at the very bottom of the hierarchy (where physical presence remains decisive) nor at the very top (where judgment and relationships dominate). The middle class of cognitive functions is on the front line.
This exposure profile has direct social implications. Economists Daron Acemoglu and Simon Johnson, in their work on the distribution of technological gains, have shown that previous waves of automation tended to compress wages for middle-class categories while providing greater benefits to the more qualified and capital owners. Nothing guarantees that agentic AI will escape this dynamic. The most exposed workers are not uniformly well-paid senior executives capable of rapid reconversion. Some of them are technicians, administrative assistants, junior analysts whose initial training does not anticipate this type of transition. This is moreover a tension that the French tax system aggravates: if capital that substitutes labor is not taxed at the same level as labor itself, the transition will occur at the expense of the workers concerned.
However, there is a counterpoint not to be overlooked. The same task-based analysis that identifies exposures also identifies emergences. AI agents create new needs: supervision of automated systems, verification of outputs, training models to sectoral specificities, design of hybrid human-machine workflows. These jobs did not exist five years ago. They do not appear in current statistics because they are being created. Arquié insists on this point: public policy must anticipate both dynamics simultaneously.
Why Companies Deploy Slowly
If capabilities are available and exposure identified, why have only one-third of large French companies launched agentic AI projects? The question is less rhetorical than it appears, because the answer directly illuminates the pace of transformation to come.
The first factor is technical. Current AI agents are powerful at bounded tasks, with defined inputs and evaluable outputs. They are far less so in environments where data is fragmented, poorly structured, or distributed across incompatible legacy systems. Yet this is everyday reality for most large French organizations. Deploying an effective agent for supplier contract management requires that contracts be stored coherently, accessible in a machine-readable format, linked to the right databases. This infrastructure is not in place in the majority of cases.
The second factor is managerial. Recomposing a position around tasks with high human added value requires that a middle manager be able to identify these tasks, redefine job descriptions, manage skills development, and absorb the transition period where productivity drops before rising again. This fine-grained management work is not equipped for. Training exists for tools; it barely exists for job reconfiguration.
The third factor is legal and social. In France, any substantial modification to a job requires dialogue with employee representatives. Projects of partial automation that affect actual employee responsibilities potentially fall within the scope of mandatory consultations. Companies that have advanced quickly have sometimes suffered significant social friction. Those advancing slowly often arbitrate in favor of social caution over technological speed. This is not necessarily a bad thing, but it is a real factor in pace.
The Real Question: Who Decides on Recomposition?
The crux of the matter is knowing who, in the organization, will have the power to define which tasks remain human and which pass to the machine, and according to what criteria.
This question is today essentially left to unilateral decisions by management. Industry or company agreements on AI use remain marginal in France. A few initiatives have attempted to set principles on human supervision of automated decisions, notably in the banking sector, but they remain isolated. Collective negotiation over work content, as opposed to wages, remains underdeveloped in French labor law.
The stakes are nonetheless considerable. A position where a significant fraction of tasks is automated can evolve in two very different ways. In the first scenario, the employer reduces headcount proportionally and maintains overall workload. In the second, the job holder retains their full-time position but focuses on tasks with high human added value, with corresponding skills development and, ultimately, better positioning in the labor market. Both scenarios are technically possible. Which becomes real depends on the balance of power in the organization, tax incentives, and the rules that collective frameworks impose or fail to impose.
This is where long-term perspective becomes concrete. By 2040, if no framework is set, the definition of what is “essentially human task” will be the result of tacit adjustments, position by position, organization by organization. The task that remains human in a bank may not be the same as in a public administration or consulting firm. The absence of conventional definition creates a risk of growing divergence in working conditions, and ultimately, in the distribution of value between capital and labor. What Diane Coyle signals about measuring progress applies here too: the tools we use to measure the economy were designed for an industrial economy. They do not capture AI productivity gains, nor the losses of workplace well-being that poorly managed recomposition can produce.
What Organizations Making Progress Are Doing
The picture is not uniformly one of waiting. Some organizations have launched experiments whose results are beginning to be documented.
In financial services, several players have deployed agents for detecting accounting anomalies and generating initial drafts of regulatory reports. Documented feedback indicates significant time savings on these tasks, and partial reorientation of teams toward analysis of complex cases and customer relations on high-stakes files. This is gradual adjustment, with its internal resistances and learning phases. But the adjustment is real.
In the public sector, the inter-ministerial digital agency is conducting experiments on using agents to process repetitive administrative requests. Published results remain preliminary, and scaling encounters the same data infrastructure problems as the private sector.
On the training side, several organizations have integrated modules on human-AI collaboration into their professional certifications. The AFPA and several OPCOs have launched specific skills development pathways for the most exposed profiles. These initiatives remain insufficient at the scale of the challenge, but they exist and are ramping up.
What these experiences share is a common characteristic: they have worked where strong intermediate management played an active role in job redefinition. Deployments piloted solely by IT departments, without involvement of field managers, have produced underutilized or poorly adapted tools. The competence for job recomposition is itself a training need.
Public Policy Has a Blind Spot
French policy on AI adaptation has focused on two axes: support for French AI supply (ecosystem funding, France 2030 plan, research programs) and regulation (active participation in the European AI Act). Both axes are useful and necessary.
But they leave a blind spot: policy on demand for recomposed work. No significant fiscal provision currently incentivizes a company to invest in skills development for an employee whose position is partially automated rather than reducing headcount. No collective bargaining mechanism has been strengthened to specifically address the task question. The 2018 reform of vocational training simplified funding but did not anticipate this type of need.
The blind spot has a simple economic logic: the costs of transition are borne by workers and the social protection system, while productivity gains are captured by companies and their shareholders. Without correcting this asymmetry, the deployment of agentic AI will be suboptimal from a social perspective, even if rational from the perspective of individual companies.
The open question is therefore not whether AI will recompose work. It is already doing so, for a growing share of workers and at a speed that will accelerate as agents become more capable. The question is whether France will give itself the political, conventional, and fiscal tools to make this recomposition happen with workers, rather than at their expense.
Sources
- Axelle Arquié, CEPII, Observatory of Threatened and Emerging Jobs, interview Nouvelle Vie Ouvrière: https://nvo.fr/axelle-arquie-les-effets-de-lia-se-feront-sentir-sur-lemploi-quand-la-technologie-sera-mature/
- Daron Acemoglu and Simon Johnson, Power and Progress (PublicAffairs, 2023), distribution of technological gains and capital-labor sharing
- Inter-ministerial Digital Agency (DINUM), experiments with AI agents in public services, annual activity reports
- Diane Coyle, GDP: A Brief but Affectionate History (Princeton University Press, 2014) and work on measuring progress in the digital age
- Report of the National Collective Bargaining Commission on the banking-retail industry agreement on AI (2023)
- Axelle Arquié, economist at CEPII and co-founder of the OEM: https://www.cepii.fr/blog/bi/contributeur.asp?auteur=Axelle+Arqui%C3%A9
- OEM - Observatory of Threatened and Emerging Jobs: https://journeeseconomieautrement.fr/intervenant/axelle-arquie/
- CEET Cnam - Company Agreements and AI (2024): https://ceet.cnam.fr/publications/connaissance-de-l-emploi/l-ia-dans-les-entreprises-que-revelent-les-accords-negocies–1501114.kjsp
- INSEE Première No. 2061 - AI Adoption in Businesses (2024): https://www.insee.fr/fr/statistiques/8604126
- AMF Report on AI Use in Financial Markets (2026): https://www.amf-france.org/sites/institutionnel/files/private/2026-02/202602-rapport-amf-ia-et-acteurs-des-marches-financiers-fr.pdf
- DINUM - Experimentation ‘The Inter-ministerial AI Assistant’: https://www.solutions-numeriques.com/la-dinum-veut-faire-de-lia-et-du-cloud-souverain-les-piliers-du-numerique-public/
- LAW No. 2018-771 - Vocational Training Reform: https://www.legifrance.gouv.fr/jorf/id/JORFTEXT000037367660
- National AI Strategy / France 2030: https://www.economie.gouv.fr/actualites/strategie-nationale-intelligence-artificielle
- Acemoglu & Johnson, ‘Power and Progress’ (2023), IMF Finance & Development: https://www.imf.org/fr/publications/fandd/issues/2023/12/rebalancing-ai-acemoglu-johnson